基于多光谱融合影像的降解膜分类与降解率估算研究

Classification of Degradation Films and Estimation of Degradation Rate Based on Multispectral Fusion Images

摘要:

为解决传统残膜污染调研，人工判别地膜耗时久、用工强度大和人为误差影响大等难题，基于无人机多光谱融合影像，采用监督分类中最大似然(Maximum likelihood classification, ML)、最小距离(Minimum distance classification, MD)和光谱角映射分类器(Spectral angle mapper classification, SAM)对棉田4种降解膜的残膜影像进行分类，并结合贝叶斯岭回归(BRR)、支持向量回归(SVR)和K近邻回归(KNNR)建模方法构建降解率估算模型，从而实现对棉田降解膜降解情况的快速调研。结果表明：ML较MD和SAM对降解膜分类效果更好，平均误差低于0.023，与实测结果相关系数均高于0.9。结合不同机器学习算法构建模型，ML-BRR降解率估算模型拟合效果和泛化能力最佳，训练集和测试集R2分别为0.756~0.966和0.823~0.921，RMSE分别不高于2.698%和3.098%。基于无人机多光谱融合影像，采用最大似然分类器进行残膜与土壤分类，并结合BRR算法构建降解率估算模型，实现对棉田降解膜降解情况快速诊断是可行的，可为残膜污染治理措施改进提供参考。

Abstract:

In order to solve the problems of traditional residual film pollution investigation, such as time-consuming manual identification of mulch film, high labor intensity and larger human error, based on UAV multispectral fusion imagery, using maximum likelihood classification (ML), minimum distance classification (MD) and spectral angle mapper classification (SAM) in supervised classification, the residual film images of four degradation films in cotton field were classified, and the degradation rate estimation model was constructed by combining Bayesian ridge regression (BRR), support vector regression (SVR) and K nearest neighbor regression (KNNR) modeling methods, so as to realize the rapid investigation of the degradation of degradation film in cotton field. The results showed that ML had a better effect on the classification of degradable films than MD and SAM, with an average error of less than 0.023 and a correlation coefficient higher than 0.9 with the measured results. Combined with different machine learning algorithms to construct the model, the ML-BRR degradation rate estimation model had the best fitting effect and generalization ability, and the R2 of the training set and testing set were 0.756~0.966 and 0.823~0.921, respectively, and RMSE were not more than 2.698% and 3.098%, respectively. Based on UAV multispectral fusion images, the maximum likelihood classifier was used to classify residual film and soil, and the degradation rate estimation model was constructed in combination with BRR algorithm, which was feasible to realize the rapid diagnosis of degradation of degradable film in cotton field, so as to provide an idea for the rapid investigation of residual film and provide reference materials for the improvement of residual film pollution control measures.